BackgroundAn estimated 30% of Major Depressive Disorder (MDD) patients exhibit resistance to conventional antidepressant treatments. Identifying reliable biomarkers of treatment-resistant depression (TRD) represents a major goal of precision psychiatry, which is hampered by the clinical and biological heterogeneity underlying MDD.MethodsTo parse heterogeneity and uncover biologically-driven subtypes of MDD, we applied an unsupervised data-driven framework to stratify 102 MDD patients on their neuroimaging signature, including extracted measures of cortical thickness, grey matter volumes, and white matter fractional anisotropy. Our novel analytical pipeline integrated different machine learning algorithms to harmonize neuroimaging data, perform data dimensionality reduction, and provide a stability-based relative clustering validation. The obtained clusters were then characterized for TRD, history of childhood trauma and different profiles of depressive symptoms.ResultsOur results indicated two different clusters of patients, differentiable with 67% of accuracy: 1) one cluster (n=59) was associated with a higher proportion of TRD compared to the other, and higher scores of energy-related depressive symptoms, history of childhood abuse and emotional neglect; this cluster showed a widespread reduction in cortical thickness and volumes, along with fractional anisotropy in the right superior fronto-occipital fasciculus, stria terminalis, and corpus callosum; 2) the second cluster (n=43) was associated with cognitive and affective depressive symptoms and thicker cortices and wider volumes compared to the other.DiscussionOur stratification of MDD patients based on structural neuroimaging identified clinically-relevant subgroups of TRD with specific symptomatic and childhood trauma profiles, which are informative for tailoring personalized and more effective interventions of treatment resistance.